Model evaluation for sfcWind

• Climate variables used as response: - sfcWind, tas, pr: original scale daily, reduced scale: hourly - clt, rsds: original scale monthly, reduced scale: daily

• Training: All models are trained using ERA5 in reduced scale as observations. Training data perdiod: 1980 - 2015.

• Validation Period: 2015-01-02 00:00:00 to 2023-12-30 23:00:00

• Response variable to predict: upscaled ERA5 (daily or monthly)

• Predictors variables: daily scale the variables sfcWind, tas, pr, tasmax, tasmin and psl and in a monthly scale clt, rsdt,rsds , the month , hour, sun’s elevation & azimuth and the daily daylight amount in seconds.

Numerical evaluation metrics

• Each row correspond an evaluation metric
• Dark blue cell indicates the best model in the row.
• Light blue cell indicates the second best model in the row.

Metric	xgboost	cnn	naive	lstm
mae	0.40	0.37	0.42	0.38
cor	0.92	0.93	0.91	0.93
ratio_of_sd	0.91	0.96	0.91	0.96
KGE	0.88	0.92	0.88	0.92
amplitude_mae	0.85	0.64	0.82	0.68
maximum_difference	5.30	4.77	5.40	4.85
sign_correlation	0.70	0.70	0.71	0.72
extreme_correlation	0.59	0.68	0.59	0.62
qqplot_mae	0.10	0.06	0.10	0.04
acf_mae	0.09	0.05	0.08	0.05
extremogram_mae	0.05	0.02	0.06	0.03

Graphical evaluation metrics

• Observations: cumulative distribution (Detrended QQ), dependency structure (ACF), and extremes (Extremogram, using 0.97 cut off).

• Daily/Monthly gruped: Maximum hour histogram (only in daily variables), Amplitude distribution.

• Hourly/Daily grouped: Mean profile per hour within days (or weeks witihin season)

• In all plots, grey color represent the true ERA5 signal.